Why Prior Sensitivity Matters
Bayesian clustering combines the observed omics data with prior distributions on mixture proportions, cluster-specific profiles, and sometimes the number of clusters. When the data strongly distinguish groups, reasonable prior changes may have little effect on posterior conclusions. When signals are weak, sparse, or discordant across modalities, the same choices can materially alter which samples are grouped together and how many clusters appear occupied. Sensitivity analysis makes that dependence visible; it does not identify a uniquely correct prior.
The target is not a label-by-label comparison of cluster names. Mixture labels can be permuted without changing the model, so a label such as “cluster 1” has no inherent identity. Instead, compare label-invariant posterior summaries, especially the probability that two samples share a cluster and the distribution of cluster sizes.
A prior specification includes more than a numerical hyperparameter: it includes the family of the prior, its support, and any structure shared across omics modalities. For example, a model might use \(z_i\mid\boldsymbol{\pi}\sim\mathrm{Categorical}(\boldsymbol{\pi})\), a prior on the cluster profile parameters \(\theta_k\), and either a fixed or random cluster count \(K\). In an integrative model, \(\theta_k\) may contain separate profile parameters for different modalities while the assignments \(z_i\) are shared.
For an interpretable comparison, keep the likelihood, preprocessing, and all non-target modeling decisions constant when changing one prior block. First check that each alternative is proper and assigns plausible probability to scientifically possible values. Prior predictive simulation is useful here: simulate profiles and allocations from the prior and ask whether their scales, sparsity, and cluster sizes are credible before examining the observed data.
Mixture-Proportion Priors and Allocation Probabilities
For a finite mixture with fixed \(K\), a common specification is \(\boldsymbol{\pi}\mid K\sim\mathrm{Dirichlet}(\alpha/K,\ldots,\alpha/K)\), where \(\alpha>0\) is the total concentration. Conditional on the assignments, conjugacy gives \(\boldsymbol{\pi}\mid z,K,\alpha\sim\mathrm{Dirichlet}(n_1+\alpha/K,\ldots,n_K+\alpha/K)\). The conditional posterior mean of \(\pi_k\) is therefore \((n_k+\alpha/K)/(n+\alpha)\). This is a statement about mixture weights given assignments; it does not by itself describe uncertainty in the assignments.
Integrating out \(\boldsymbol{\pi}\) also shows directly how this prior affects allocations. For a particular assignment vector \(z\), with counts \(n_1,\ldots,n_K\), its prior probability is
This follows by multiplying the categorical assignment probability \(\prod_k\pi_k^{n_k}\) by the Dirichlet density and integrating over the simplex. The result makes clear why changing \(\alpha\) can alter the relative prior support for balanced and unbalanced allocations. The likelihood also matters: a prior preference is not a guarantee that the posterior will favor the same allocation.
Worked Example: Conditional Mixture-Weight Sensitivity
Suppose, for illustration, that \(n=12\), \(K=2\), and a particular allocation has counts \(n_1=9\) and \(n_2=3\). Compare total concentrations \(\alpha=2\) and \(\alpha=20\). These calculations condition on this allocation; they are not a substitute for comparing the full posterior over allocations.
With \(\alpha=2\), the Dirichlet parameters are \(9+1=10\) and \(3+1=4\). Thus
With \(\alpha=20\), each component receives prior parameter \(10\), so
For these fixed assignments, the larger concentration pulls the expected proportions toward equality. It does not change the observed counts of 9 and 3, and it does not establish how often those assignments occur after integrating over uncertainty in \(z\). A full sensitivity analysis must compare posterior draws of the assignments as well as conditional weight summaries.
Comparing Posterior Co-Clustering and Cluster Sizes
For each prior specification \(r\), estimate \(C_{ij}^{(r)}\) from posterior draws. If \(S\) draws are available, the Monte Carlo estimate is the fraction of draws in which the pair shares a cluster. Since this event is unchanged by relabeling, it remains meaningful even when cluster labels switch. For cluster sizes, summarize the posterior distribution of the sorted size vector, or report the distribution of the number of occupied clusters and selected size summaries. Sorting the sizes avoids treating arbitrary labels as identities.
Differences such as \(\Delta_{ij}=C_{ij}^{(A)}-C_{ij}^{(B)}\) describe how much a particular pair’s posterior co-clustering probability changes between two priors. A large difference may matter substantively, but it should be considered alongside Monte Carlo error and the broader pattern across pairs. For example, a heat map of the full co-clustering matrix can reveal whether a change is localized or reorganizes many samples.
Worked Example: Profile-Prior Sensitivity in Co-Clustering
Consider two plausible prior scales for cluster-specific profiles in a hypothetical integrative model. Prior A allows more variation among cluster profiles; Prior B concentrates profiles more tightly around their prior centers. Suppose two independent posterior runs, each summarized by 1,000 independent draws for this arithmetic illustration, give the following same-cluster counts for three pairs.
| Sample pair | Prior A: same cluster | Prior B: same cluster |
|---|---|---|
| 1 and 2 | 920 of 1,000 | 610 of 1,000 |
| 3 and 4 | 840 of 1,000 | 790 of 1,000 |
| 5 and 6 | 480 of 1,000 | 510 of 1,000 |
For samples 1 and 2, the estimated co-clustering probabilities are \(920/1000=0.920\) and \(610/1000=0.610\). Their difference, Prior A minus Prior B, is \(0.920-0.610=0.310\). For samples 3 and 4, the corresponding estimates are \(0.840\) and \(0.790\), a difference of \(0.050\). For samples 5 and 6, the estimates are \(0.480\) and \(0.510\), a difference of \(-0.030\).
In this hypothetical output, the first pair’s grouping is notably more sensitive to the profile prior than the other two pairs. The estimates do not show that Prior A is better, nor do they establish a cause for the difference; the data and prior jointly determine the posterior. A complete analysis would inspect all pairs and the posterior distribution of sorted cluster sizes under both priors, then check that the difference is larger than simulation error.
Prior changes can influence one another. A profile prior affects how well a proposed grouping can explain the observed features; a mixture-proportion prior affects allocation probabilities; and a prior on \(K\) affects which model dimensions are available or favored. One-at-a-time comparisons help identify which block is influential, but they can miss interactions. A final set of joint, coherent scenarios is useful when several prior assumptions are tied to the same scientific judgment.
Cluster-Count Priors and a Posterior-Odds Justification
When \(K\) is unknown, compare alternatives that put plausible prior probability on different values of \(K\), or use an appropriate nonparametric mixture specification. Be explicit about whether a reported count is the nominal finite-mixture dimension or the number of occupied clusters. In a finite mixture, some components can be empty, so those counts need not agree. For an unknown-\(K\) model whose \(K\) denotes occupied clusters, state that convention and summarize the posterior distribution of that count.
The reason prior sensitivity is expected can be seen from posterior odds. Let \(z_A\) and \(z_B\) be two candidate allocations under the same model specification. Integrate the cluster parameters against their prior distributions to obtain \(m(Y\mid z)\), the marginal likelihood for the data conditional on an allocation. Bayes’ rule gives
The first ratio measures the relative data support after accounting for parameter uncertainty; the second is the prior allocation odds. If a prior changes, the second ratio can change, and a cluster-profile prior can also change the integrated likelihood ratio. Thus posterior grouping can move even when \(Y\) is unchanged. This identity explains the mechanism; it does not imply that every prior change must produce a large posterior change.
Worked Example: Sensitivity to a Prior on Cluster Count
Suppose two plausible priors on the number of occupied clusters are compared for the same hypothetical data and likelihood. In 200 independent posterior draws under Prior A, the occupied-cluster counts are \(K=2\) in 20 draws, \(K=3\) in 120 draws, and \(K=4\) in 60 draws. Under Prior B, they are \(K=2\) in 100 draws, \(K=3\) in 80 draws, and \(K=4\) in 20 draws.
The estimated posterior probabilities under Prior A are \(20/200=0.100\), \(120/200=0.600\), and \(60/200=0.300\). Its posterior mean count is
Under Prior B, the estimated probabilities are \(100/200=0.500\), \(80/200=0.400\), and \(20/200=0.100\), and the posterior mean is
The posterior mean differs by \(3.20-2.60=0.60\) occupied clusters, and the posterior probability of exactly two clusters differs by \(0.500-0.100=0.400\). These are substantial differences in this hypothetical example. They should be reported with the posterior distributions, not only the means, and interpreted as sensitivity of the clustering conclusion to the stated prior alternatives.
Practical Workflow, Diagnostics, and Alternatives
A defensible sensitivity analysis follows a controlled sequence. Convergence and adequate exploration matter: a chain that has not explored its posterior can mimic prior sensitivity or conceal it. Use the convergence diagnostics for clustered omics chains, including multiple chains and checks on label-invariant summaries. For each prior comparison, use sufficient effective draws and assess Monte Carlo uncertainty in the quantities being compared. Raw retained draw counts alone do not guarantee precise estimates when draws are autocorrelated.
State the changed prior family or hyperparameter and why each value is scientifically reasonable. Check prior predictive implications for cluster sizes and profile scales.
Use the same data, likelihood, preprocessing, and computational settings when isolating one prior block. Record any additional changes in joint scenarios.
Compare co-clustering probabilities, occupied-cluster counts, and sorted cluster-size distributions rather than raw label names.
Assess Monte Carlo precision and chain mixing, then identify which sample relationships or size conclusions remain stable and which depend on the prior.
One-at-a-time analysis is an alternative to changing every prior simultaneously: it is easier to attribute differences, but it may understate interactions. Joint scenario analysis captures combinations of assumptions but makes attribution harder. A broad range of hyperparameters is not automatically informative; alternatives should be plausible, and conclusions should not be based on implausible extremes. Posterior predictive checks address model-data fit, while prior sensitivity addresses dependence on prior choices; neither replaces the other.
Common Pitfalls and Limitations
- Comparing cluster labels directly. Label switching can make “cluster 1” in one fit correspond to a different group in another. Compare co-clustering and sorted sizes instead.
- Calling a prior effect a data discovery. If a grouping changes under plausible priors, report that dependence rather than presenting one result as uniquely determined by the data.
- Confusing Monte Carlo noise with sensitivity. Small apparent changes may be simulation error. Check chain behavior and precision for the specific posterior summaries.
- Changing several components without saying so. A simultaneous change to the proportion, profile, and cluster-count priors cannot reveal which assumption produced the difference.
- Equating nominal and occupied cluster counts. Empty components in a finite mixture mean the two counts can differ; state which one is summarized.
Sensitivity analysis has limits. It can show that conclusions are robust across the alternatives examined, but it cannot prove robustness to every possible prior. It also cannot repair a misspecified likelihood, poor convergence, or unmodeled dependence among modalities. Report the alternatives, the summaries that changed, and the conclusions that remained stable.
Check Your Understanding
Answer each question, then compare your reasoning with the solution immediately below it.
- A pair of samples shares a cluster in 735 of 900 independent posterior draws under one prior. What is the estimated co-clustering probability?
Solution: The estimate is \(735/900=0.8167\), or about 0.817. It estimates the posterior probability of shared assignment under that prior and model, not a probability that the prior is correct.
- In a two-component mixture, assignments have counts 7 and 5, and the total concentration is \(\alpha=4\). What are the conditional posterior means of the two mixture proportions?
Solution: Each Dirichlet prior parameter is \(\alpha/K=2\). The conditional posterior parameters are 9 and 7, so the means are \(9/(9+7)=0.5625\) and \(7/16=0.4375\).
- Why is the number of occupied components not necessarily the same as the nominal \(K\) in a finite mixture?
Solution: A component is occupied only if at least one sample is assigned to it. Some of the \(K\) available components can have zero assignments, so the occupied count can be less than \(K\).
- Two priors produce different posterior co-clustering estimates, but one chain has poor mixing. Is this sufficient evidence of prior sensitivity?
Solution: No. Poor mixing makes the posterior estimates unreliable and can create an apparent difference. Diagnose and improve computation, then compare label-invariant summaries with adequate Monte Carlo precision.